链接：https://pan.baidu.com/s/1weaiBh0_yH2BQni5eTxHgg 提取码：qvsq
链接：https://pan.baidu.com/s/1FClmX0SH3WarcczkEQbmwA 提取码：ps8j
链接：https://pan.baidu.com/s/1zaR6aXh9PVQs063Q_b9zQg 提取码：ajma
链接：https://pan.baidu.com/s/1UksOGtD2yl6k8mtB-Wr39A 提取码：4f68
链接：https://pan.baidu.com/s/1KHMlbhu5R29CJvundGL4Sw 提取码：8wtc
链接：https://pan.baidu.com/s/1iVVAXS4LZLvoQSGQnz7ROw 提取码：d53m
Here, we have only uploaded the result file of the paper (PR_0.777 SR_0.539 on RGBT234, PR_0.894 SR_0.724 on GTOT.)
This code is an updated version, simplified from the one we submitted for the VOT2019-RGBT challenge.
Consequently, there are some differences compared to MANet's paper.
CPU: Intel(R) Core(TM) i7-7700K CPU @ 3.75GHz GPU: NVIDIA GTX1080（8GB） Ubuntu 16.04
- python2.7
- pytorch == 0.3.1
- numpy
- PIL
- some others library functions
In our tracker, we use MDNet as our backbone and extend it to a multi-modal tracker.
We use imagenet-vgg-m.mat as our pretrain model.
You can choose either a two stage training or end2end training
- Stage1: Use the RGBT dataset to train whole network, then save the final model;
- Stage2: Load only the parameters of GA from the model saved in Stage1, and use the same RGBT dataset to train MA and IA while keeping GA fixed.
- The training method here is the same as with MDNet
Pretrain model :https://drive.google.com/open?id=1aO6LhOTxmpd7o_JXPLPjL3LsrQ5oqbl7
In the tracking/run_tracker.py file, you need to change the dataset path and save the result file directory. In the tracking/options.py file, you need to set model file path and set learning rate depend on annotation. For the testing and training stages, update the 'modules/MANet3x1x1_IC.py' file depending on the annotation.
Tracking model:https://drive.google.com/open?id=1Png508G4kQPI6HNewKQ4cfS36CvoSFSN
